The AI assistant is built into the web interface of every Weble IoT Universal Gateway. Open it with the AI button in the header and talk to your device in plain language: ask what a gateway is doing, find why a route stopped, read logs, create addresses and routes, diagnose the network, back up the configuration, or write a maintenance script. The assistant works through the same API and the same access rights as the interface, and asks for your permission before touching anything.
It runs in one of three modes:
Assistants can also run unattended on the device and answer over SMS, WhatsApp, MQTT or any other protocol: see the AI driver.
The language model runs in the Weble cloud and at the chosen AI provider, never on the device. Your messages and the information the assistant consults are sent there. See section 8.
Devices installed before version 9.1 (r722, 27 October 2026) do not have credits yet. The Credits window shows how to request the 5 $ welcome credit; devices prepared since then receive it at installation.
The Built-in mode is paid with AI credits held on a balance attached to the device. One credit is worth 0.01 $. Each answer shows the number of tokens used and its cost in credits. The price depends on the model, as shown by the cost meter in the model menu.
The Credits window (model menu, Connection, Manage) shows:
Credits are ordered on Weble.ch or from support@Weble.ch. When the balance reaches zero, the assistant refuses new turns until credits are added; conversations and permissions are kept.
A credit licence is a shared balance identified by a code of the form WL-XXXX-XXXX-XXXX. Attach it to a device from the assistant's Credits window or from Administration › Licences (Insert licence recognises the WL format). Several licences can be attached to one device, with a priority order.
The code can be entered while the device is offline: it is stored and attached at the next connection to the AI service, by whoever opens the assistant. The Licences page then shows the licence as a regular row with its remaining balance, its last refresh date and a refresh button.
Attaching, detaching or transferring licences requires an administrator account, or the right to insert product licences.
The chat on credits needs no product licence. Two features require the AI Connect licence (product code 1010) or the FULL licence (2628, 9779):
| Feature | Without licence | With AI Connect or FULL |
|---|---|---|
| AI driver | does not start | enabled |
| MCP connector | tokens cannot be created on the device | enabled |
| Your own API key | allowed, with a service retention of 20 % of the credit equivalent of each message | no retention |
The model menu groups the available models by provider, with a cost, speed and capability meter and a note on data residency for each one. The list is maintained server-side and grows without a device update.
| Provider | Models | Notes |
|---|---|---|
| Anthropic | Claude Haiku (default), Claude Sonnet, Claude Opus, Claude Fable | Reads files and images. Reasoning effort adjustable on Sonnet, Opus and Fable; extended thinking on Haiku, Sonnet and Opus. Fable costs twice Opus. |
| OpenAI | GPT-5.5, GPT-5.4 mini, GPT-6 Astra | Reasoning effort adjustable. Astra does not read attached images and costs twice GPT-5.5. |
| DeepSeek | DeepSeek fast, DeepSeek pro | Data processed on servers in China: avoid for sensitive data. Cannot read files or images. |
The model is chosen per conversation; switching in the middle of a conversation shows a warning. Your last choice is remembered in the browser for this device.
Choose My key in the model menu and add a key for Anthropic, OpenAI or DeepSeek. The key is tested with a minimal call, then stored encrypted on the Weble server for your account on this device. Only a masked form is shown afterwards; the key never appears in conversations or logs. You can save one key per provider and remove them at any time.
In key mode the model list is limited to the key's provider. If My key is selected and no key is saved, the input is blocked rather than silently charging credits.
The assistant talks to the device through its peers API: gateways, addresses, routes, loggers, logic, alarms, system, network, accounts, views. Anything the interface can do, the assistant can do, within your rights and with your permission. On top of the API it has dedicated tools:
Reading and analysing
Acting
Files
The assistant's browser
Once you allow it, the assistant can open the device's own web pages in a headless browser, logged in with your account and your rights: open a page, take a screenshot, click, fill a form, read the page text. Every browser session is announced by a card and shown by a header pill, Assistant browser active, with a Close button that ends it immediately. The browser session ends with yours.
The assistant starts locked: even reading needs your approval. Permissions are organised in groups ordered by risk:
| Group | Scope | Risk |
|---|---|---|
| Read | devices, values, logs, configuration | — |
| Network & field diagnostics | ping, traceroute, interfaces | low |
| Process logs | logs of the device's processes | low |
| Write values | write address values | low |
| Add | create gateways, addresses, routes, logic… | low |
| Edit | modify, enable, disable, save | medium |
| Scripts | run scripts in the sandbox | medium |
| Delete | delete, purge, reset | high |
| Services | restart, stop, reboot | high |
| Accounts | accounts, secrets, certificates | critical |
When the assistant needs something it does not have, a permission card offers: this session, whole group; this session, this function only for the risky groups; just this time; or deny. Session grants are shared by all your windows on the device and expire one hour after the last window closes. The permissions button shows a colour pill, grey for none, green for read-only, orange for modification, red for deletion, services or accounts, and lets you allow everything for the session or revoke all.
Three rules always apply:
There is no permanent grant for the interactive assistant: every session starts locked. Permanent grants exist only for AI driver assistants, which must work unattended.
Conversations are stored per user on the Weble server, not in the browser: you find them again from another computer, and an account never sees another account's conversations. The panel lists them and lets you rename, delete and search them. Memory is per conversation; each thread keeps its own context.
The panel follows the interface language, shows token counters per message and per conversation, and a Stop button interrupts a running answer. A turn is limited to 25 tool calls.
The device opens an outgoing, certificate-verified connection to the Weble AI service and identifies itself with its own token. Nothing needs to be opened in your firewall.
If the device has no internet access but your computer has, the panel can connect through your browser: the browser reaches the AI service and relays it to the device. The device still authenticates itself; the browser only carries the traffic. When the AI service does not answer, the device waits longer between attempts rather than retrying every few seconds.
| Page | Description |
|---|---|
| MCP: external AI clients | Connect Claude Desktop or any MCP client to the device, tokens, takeover, exposed tools. |
| AI driver | Assistants running on the device, triggered by routes, answering over SMS, WhatsApp, MQTT. |
| Releases | Version history: r718, r720, r721, r722 (v9.1), r723. |